Vulnerability to climate change of islands worldwide and its impact on the tree of life

Island systems are among the most vulnerable to climate change, which is predicted to induce shifts in temperature, rainfall and/or sea levels. Our aim was: (i) to map the relative vulnerability of islands to each of these threats from climate change on a worldwide scale; (ii) to estimate how island vulnerability would impact phylogenetic diversity. We focused on monocotyledons, a major group of flowering plants that includes taxa of important economic value such as palms, grasses, bananas, taro. Islands that were vulnerable to climate change were found at all latitudes, e.g. in Australia, Indonesia, the Caribbean, Pacific countries, the United States, although they were more common near the equator. The loss of highly vulnerable islands would lead to relatively low absolute loss of plant phylogenetic diversity. However, these losses tended to be higher than expected by chance alone even in some highly vulnerable insular systems. This suggests the possible collapse of deep and long branches in vulnerable islands. Measuring the vulnerability of each island is a first step towards a risk analysis to identify where the impacts of climate change are the most likely and what may be their consequences on biodiversity.


Supplementary Method S1: Correction of expected diversity loss values
Summary a. General

a) General methodology to correct expected diversity loss values
GBIF data are biased by the uneven geographic coverage among islands We included this source of uncertainties using a two-step procedure. First, we calculated an index of geographic coverage for each island. To do so, we estimated a prediction of genus richness in each island from a Boosted Regression Trees model (hereafter "BRT"). In this model, the response variable is the number of genera in an island -calculated from the GBIF -and the explanatory variables are a combination of bioclimatic, geographic and historic variables (see Supplementary Method S1(b)). BRT is an ensemble method for fitting statistical models that combines algorithms of regression trees and boosting 1 . We chose the optimum number of trees based on learning rate, tree complexity and bag fraction that we then used to predict values of species richness. We finally calculated an index of geographic coverage for each island as the observed value of species richness over the predicted valued.
In a second step, we performed linear models of ExpPDloss (and ExpSRloss and ExpSRlossED [Supplementary Method S2]) in function of our index of coverage. Corrected values of ExpPDloss, (and ExpSRloss and ExpSRlossED) were defined as the residuals of this linear model. We performed calculations both to correct losses estimated from all genera found in islands and for genera strictly endemic to islands (no occurrence on continents).

b) Variables tested
To predict values of species richness we tested the effect of geographic, bioclimatic and historical factors: Geographical and physical factors of islands: area (km²), elevation (meters), minimum distance to continent (km), the proportion of surrounding landmass, latitude and longitude. Species diversity may theoretically increase with area and proximity to continent and these factors were both shown to be among the strongest predictor of plant species richness 2 . As for the proportion of surrounding landmass (SLMP), which accounts for the size and coastline shape of surrounding landmass, it has been found to be a valuable isolation metric to explain island plant diversity at a global scale 3 . Maximal elevation is related to topography and environmental heterogeneity, for example due to temperature decrease with altitude 4 . This was shown to explain a great proportion of insular plant species richness 2 . As for latitude and longitude, they may highlight differences related to the geographic position of islands and especially the diversity latitudinal gradients 5 .
ii. Bioclimatic factors: we selected the number of ecoregions occurring in an island, mean annual temperature (C°), mean annual rainfall (mm), temperature seasonality (C°), rainfall seasonality (mm), mean and standard deviation in annual solar radiation (kJ m -2 day -1 ), mean and standard deviation in annual wind speed (m s -1 ), mean and standard deviation in annual water vapor pressure (kPa), isotherm (C°).
Bioclimatic factors may filter out lineages weakly adapted to the environmental conditions of an area 6 . Especially, temperature and rainfall are among the main predictors of plant species richness 2 , 7 , 8 . Wind speed may act on plant diversity by favoring long distance dispersal 9 , influence plant growth and may select for lineages adapted to cool or harsh wind conditions. Evapotranspiration, measured from water vapor pressure, and solar radiation are key components related to energy availability. They were both shown to be predictors of plant richness in islands as well as in continents 7 , 9 . As for the number of ecoregions it reflects areas with distinct environmental conditions and distinct assemblage of natural communities sharing a large majority of species 10 . Temperature, rainfall, wind speed, solar radiation and vapor pressure data were extracted from Worlclim version 2 11 and missing data were completed thanks to the GID database 12 . We used the ecoregions defined by the WWF 10 .
iii. Historical factors. We used the velocity of past climate change.
The effect of past climate change velocity may depend on the taxon considered but low velocity were generally associated to high rates of endemism at regional scales 13 . Data of velocity of past climate change came from Sandel et al. 13 .
iv. Sampling effort: we tested for the effect of the ICEr metric described below in Supplementary Method S1 (c).
To avoid collinearity in the models, the following variables, showing high correlations (Pearson correlation coefficients > 0.5), were removed from models: temperature seasonality, isotherm, mean annual solar radiation, standard deviation in solar radiation, mean annual vapor pressure and standard deviation in wind speed.

c) Measuring sampling effort
To estimate sampling effort, we chose the Incidence-based Coverage Estimator (ICE ; 14 ) which is estimated from the number of rare species in sub-samples and species accumulation curves. Compared to other estimators of species richness, the ICE may best satisfy the requirements for an ideal species-richness estimator such as the capacity to reach a stable value independently of sample size, the low sensitivity to sampling patchiness and density. To calculate ICE, we defined a sub-sample in each island as a set of observations obtained at a given date. We then estimated ICE thanks to the R function spp.est found in package 'Fossil' 15 . Finally, we calculated the ratio of the observed number of species in an area on the expected number of species estimated from ICE, an index we called ICEr.

d) Contribution of variables to species richness
To reach convergence, we parameterized BRT with a bag fraction of 0.5, a learning rate of 0.005 and a tree complexity of 8,700 trees. The contribution of each variable is represented in Figure S1 Figure S1: Contributions of biotic variables to genus richness in world islands

Supplementary Method S2: Expected loss of species richness and evolutionary distinctiveness
We calculated the expected loss of SR and its significance as the sum of plant extinction probabilities following the loss of an island.

( ) = ∑
where i designates the i th species (monocot genera in our study) and pi denotes its extinction probability.
We assessed the loss of species richness in the most evolutionary distinct species (SRED) as

Global vulnerability
All genera Endemic genera